Daily Paper Briefing

Abstract Overview

Chinese-Jev is a Chinese-language System One model designed for bounded decision tasks, including option selection, proposition judgment, and ordinal scoring. The authors develop a unified pipeline that converts heterogeneous Chinese annotations into probability targets over candidate options, enabling a shared training formulation across general and domain-specific tasks. Built on a lightweight encoder-only backbone with decision-oriented candidate scoring, the model undergoes broad Chinese pre-training on a 10 million-example general corpus followed by independent fine-tuning for medical, legal, and financial domains. The paper also presents CJ-Bench, a held-out benchmark containing 307,900 decisions across general and specialized Chinese settings to evaluate accuracy, calibration, and latency.

Novelty

The paper adapts the System One decision paradigm to Chinese tasks via a unified supervision format that translates diverse annotations into candidate-probability targets across choice, noul, and score tasks. This is paired with an efficient encoder-only candidate-scoring architecture, a two-stage general-to-specialist training pipeline, and a dedicated multi-domain evaluation benchmark.

Results

On the general subset of CJ-Bench, Chinese-Jev General achieves 69.20% accuracy and 3.78% ECE, surpassing the hosted Jev API in accuracy with a 20.3× lower measured latency under the reported setup. Domain-specific fine-tuning increases average specialized domain accuracy from 37.97% to 70.82%, with the medical specialist reaching 87.65% accuracy (exceeding Jev by 4.0% relative) and specialists averaging 92.0% of Jev's accuracy at 15 ms median latency. Additionally, an INT8-quantized browser deployment demonstrates local execution on an iPhone 15 Pro at approximately 1 second per decision.

Key Points

1. A unified data construction pipeline converts heterogeneous Chinese annotations into probability distributions over candidate options for choice, noul, and score decision types.
2. The authors introduce CJ-Bench, a 307,900-decision benchmark covering general, medical, legal, and financial tasks to evaluate decision accuracy, expected calibration error, and inference latency.
3. Chinese-Jev achieves competitive accuracy and improved calibration compared to prior baselines at low inference latency, and it supports local on-device mobile execution via INT8 quantization.

References

• arXiv: https://arxiv.org/abs/2609.36965v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2609.36965v1
• Hugging Face Papers: https://huggingface.co/papers/2609.36965
• Project: https://gulucaptain.github.io/Chinese-Jev/